Definition
A security-policy gap where the system compares a command's name differently than the enforcement engine does, so an action the operator denied slips through. The agent ends up executing an operation the policy intended to block or flag. It defeats a well-intentioned 'least privilege' policy for an AI agent.
Why it matters
It shows that a carefully scoped AI-agent permission policy can be quietly circumvented, undermining intended control and auditability.